groxaxo/google-latam-spanish-boundary-normalized

Google LATAM Spanish Boundary-Normalized Audio

Female Spanish speech from the following upstream datasets:

• —Argentina: `ylacombe/google-argentinian-spanish`
• —Chile: `ylacombe/google-chilean-spanish`
• —Colombia: `ylacombe/google-colombian-spanish`

Attribution and Thanks

Many thanks to [ylacombe](https://huggingface.co/ylacombe) for publishing and maintaining the original Argentinian, Chilean, and Colombian Spanish datasets. The recordings, transcripts, speaker labels, and original dataset structure come from that work.

This derivative does not alter the spoken content or transcripts. Our changes are limited to:

• —detecting existing leading and trailing boundary audio with 10 ms RMS frames at -40 dBFS;
• —trimming those detected boundaries;
• —inserting exactly 0.500 seconds of digital silence before and after speech;
• —updating each manifest's duration field to match the resulting WAV; and
• —adding a reproducible processing report and dataset documentation.

Processing

Leading and trailing boundary audio was detected with 10 ms RMS frames at -40 dBFS. Existing boundary silence was trimmed, then exactly 0.500 seconds of digital silence was added at both the beginning and end of every WAV.

The operation preserved each file's 48 kHz sample rate and PCM subtype. The normalizer uses atomic replacement and is idempotent.

Contents

Each country contains a female/manifest.jsonl and a female/wavs/ directory. Manifest rows include the transcript, speaker, source dataset, source index, audio path, and normalized duration.

boundary_silence_normalization.json contains the complete per-file before/after processing report.

License

CC BY-SA 4.0, matching the upstream dataset declarations. Preserve upstream attribution when redistributing or adapting this dataset.